Abstract: Hyperspectral image classification is an challenging problem in remote sensing. Hyperspectral image classification based on deep learning framework has attracted more and more attention due to its excellent classification performance. However, the common drawback of these methods is that the training of the model requires not only a lot of time, but also a large number of labeled samples. To address this issue, a superpixel graph convolution network-based hyperspectral image classification method is proposed. The method takes superpixels instead of pixels as nodes of graphs, which greatly reduces the graph size and improves the classification efficiency. The proposed superpixel merging and smoothing techniques effectively fuse spectral and spatial information and enhance the role of spatial information in classification. To validate the effectiveness of the proposed method, experiments are carried out on two real datasets, Indian Pines and Pavia University. Furthermore, the method is also compared with some advanced hyperspectral image classification methods based on deep learning framework. The results show that the proposed method is superior to other methods in classification accuracy and efficiency.

Key words: hyperspectral image, graph convolution network, superpixel, classification